Topic
Suites that pass because they never exercise the production entry point, including self-referential tests that carry their own stale definition of valid.
Unit tests prove a function behaves when something calls it, and nothing else. Proving that production still calls it needs a check that discovers the handlers independently of the dispatch list it audits.
Reality
One developer disabled each fail() call in his own engine and re-ran the suite once per line. The 36/36 green baseline measured whether gates bite on dirty input rather than whether any test reaches them.
Reality
story-cli's MCP server answered nothing in a real repo while its protocol tests passed. The fault was process.exit in the CLI entry point, and no unit test owns a process.
Reality
WhyRep bounded its plateau read to 14 misses plus two rows. Any lifter who deloaded once during a stall stopped being told they had plateaued, and nothing logged it.
Reality
A developer set Dynamic Type to the largest size, screenshotted every screen, and surfaced the same layout defect eleven times. In Japanese it fails much earlier than in English.
Reality
A seventh picture-book theme was added in nine places and failed at one: a six-literal Zod enum in a route handler that no test path ever called.
Reality